20100095284 | CACHING RUNTIME GENERATED CODE | April, 2010 | Herring et al. |
20080263518 | SOFTWARE DOCUMENTATION MANAGER | October, 2008 | Bank et al. |
20060282818 | Interactive formula builder | December, 2006 | Despain |
20080141233 | Presenting machine instructions in a machine-independent tree form suitable for post-link optimizations | June, 2008 | Gurevich et al. |
20080066055 | Reversing execution of instructions in a debugger | March, 2008 | Shebs |
20060053418 | Network management system configuring | March, 2006 | Lundstrom et al. |
20050187756 | System and apparatus for handling presentation language messages | August, 2005 | Montgomery et al. |
20080313608 | SYSTEM AND METHOD FOR VARIABLE TYPE IDENTIFICATION | December, 2008 | Cowtan |
20080052663 | PROJECT EXTENSIBILITY AND CERTIFICATION FOR STACKING AND SUPPORT TOOL | February, 2008 | Cope et al. |
20040015887 | Simulation method, simulation system and simulation program for software and hardware | January, 2004 | Kudo |
20080263505 | Automated management of software requirements verification | October, 2008 | Stclair et al. |
IBM® is a registered trademark of International Business Machines Corporation, Armonk, N.Y., U.S.A. Other names used herein may be registered trademarks, trademarks or product names of International Business Machines Corporation or other companies.
1. Field of the Invention
This invention relates to reductions of array variables, and particularly to a method for solving a general reduction problem, which is generalized as a divide-and-conquer problem for solving a problem on an instance of size n, by dividing the instance into two or more smaller instances.
2. Description of Background
Most compilers are capable of recognizing simple reductions such as the accumulation into the variable “SUM” 10 in FIG. 1. Reductions of array variables are also common in scientific codes and are a potential source of significant improvements in parallelization results. In general, a reduction can be expressed in the form of R={r,f,d}, where r can be a scalar variable, an array, or a compound expression. The symbol, represents an associative reduction operator, is a function of reduction execution, and d is a domain for reduction operation. Taking the code segment 10 in FIG. 1 as an example, r is a scalar variable, i.e., SUM. is +, f is an identity function, and d is the data area pointed by A ranging from A[1,1] to A[N,M].
Handling such kind of scalar reduction in FIG. 1 is straightforward. There are a number of solutions implemented or proposed. For example, a conventional method is to use a critical section to enable parallel execution of the loop. Another approach is to compute a partial reduction result of each processor and then merge the results to get the overall result. However, for a reduction that involves complicated structures, such as a conditional branch, the identification and implementation is difficult.
For instance, an example code segment 12 in FIG. 2 is considered, which is used to calculate a global maximum and indexes associated with the maximum value. The loop in FIG. 2 cannot be parallelized due to the loop carried dependences on Max_value, Max_index_i and Max_index_j. Another example of a code segment 14 is shown in FIG. 3, which is used to detect the number of array elements whose value is equal to that of variable val. Code segment 12 in FIG. 2 and code segment 14 in FIG. 3 illustrate the difficulty in identifying and implementing a general reduction.
Considering the limitations of the aforementioned methods, it is clear that there is a need for an efficient method for solving a general reduction problem, which is generalized as a divide-and-conquer problem for solving a problem on an instance of size n, by dividing the instance into two or more smaller instances.
The shortcomings of the prior art are overcome and additional advantages are provided through the provision of a method for providing parallel processing capabilities, the method comprising: performing scalar and array privatization analysis via a compiler; checking whether an assignment statement is reducible; recognizing reduction patterns through a pattern matching algorithm; classifying a reduction type of each of the reduction patterns; and performing transformations and code generation for each reduction the reduction type of each of the reduction patterns.
The shortcomings of the prior art are overcome and additional advantages are provided through the provision of a system for providing parallel processing capabilities, the system comprising: a network; and a host system in communication with the network, the host system including software to implement a method comprising: performing scalar and array privatization analysis via a compiler; checking whether an assignment statement is reducible; recognizing reduction patterns through a pattern matching algorithm; classifying a reduction type of each of the reduction patterns; and performing transformations and code generation for each reduction the reduction type of each of the reduction patterns.
Additional features and advantages are realized through the techniques of the present invention. Other embodiments and aspects of the invention are described in detail herein and are considered a part of the claimed invention. For a better understanding of the invention with advantages and features, refer to the description and the drawings.
As a result of the summarized invention, technically we have achieved a solution that provides for an efficient method for solving a general reduction problem, which is generalized as a divide-and-conquer problem for solving a problem on an instance of size n, by dividing the instance into two or more smaller instances.
The subject matter, which is regarded as the invention, is particularly pointed out and distinctly claimed in the claims at the conclusion of the specification. The foregoing and other objects, features, and advantages of the invention are apparent from the following detailed description taken in conjunction with the accompanying drawings in which:
FIG. 1 illustrates one example of a reduction variable code segment;
FIG. 2 illustrates one example of a maxloc loop code segment;
FIG. 3 illustrates one example of a reduction loop code segment;
FIG. 4 illustrates one example of a code segment for partial summation according to the exemplary embodiments of the present invention; and
FIG. 5 illustrates one example of a code segment summing up the partial result according to the exemplary embodiments of the present invention.
One aspect of the exemplary embodiments is a method for solving a general reduction problem, which is generalized as a divide-and-conquer problem for solving a problem on an instance of size n, by dividing the instance into two or more smaller instances. In another exemplary embodiment, the outermost loop is parallelized, which results in a performance improvement on multi-processor systems.
The exemplary embodiments of the present invention present a framework to solve a general reduction problem, which is generalized as a divide-and-conquer problem. In particular, solving a problem on an instance of size n, by dividing it into two or more smaller instances. Each of these smaller instances is recursively solved, and the solutions are combined to produce a solution for the original instance. The exemplary embodiments consist of scalar and array privatization analysis, general reduction recognition and type classification, parallelization, and reduction coalescing and aggregation.
Concerning scalar and array privatization on analysis a compiler performs scalar and array privatization analysis. The compiler finds those scalar variables whose definitions and references are closed in a loop iteration. Furthermore a single array data-flow analysis is used to determine arrays involved in data dependences, to locate private arrays and to recognize reductions. Array data-flow analysis is a bottom-up inter-procedural analysis on the loops and procedures of the program, using the region-based analysis framework.
Concerning general reduction detection through pattern matching, the exemplary embodiments use a pattern matching algorithm to identify the MAXLOC/MINLOC reduction using the form of an “if structure”. The absolute value operations are fully supported. The exemplary embodiments start from the innermost loop. Taking the code segment 12 in FIG. 2 as an example, since it satisfies the following conditions, it would be considered as a MAXLOC reduction. The conditions are as follows: (1) the loop body starts with an if statement, S1, (2) there are a total of three statements inside the if structure, which is one more than the dimension of the array, (3) there is a maximum value assignment, S2, the array element is exactly the same as that of the if statement, (4) for the location assignment, S3 and S4, the right side is the index of the array, and (5) after the reduction in the innermost loop is identified, propagation to the outer-most loop in accordance with the dimension of the array is performed.
Concerning the registering of the reduction, the reduction is added to the reduction list of the top-most nesting level. For extreme reductions, a reduction set may be required. Taking the code segment 12 in FIG. 2 as an example, the code segment 12 includes the variable Max_value, which is used to record the maximum value of the array A along with Max_index_i and Max_index_j to recall the position of the maximum value. For the convenience of code generation, the exemplary embodiments use two reduction types: MAXVAL/MINVAL, which is used for extreme values and MAXLOC/MINLOC, which is used for the location of the extreme values. The relationship of the set of reductions can be expressed as follows:
in which u denotes the array element and i,j . . . denote the corresponding indices.
In case the two elements are equal, the index is kept unchanged, which is compatible with sequential logic.
Concerning parallel reduction code generation, a parallelizer generates three steps for each reduction variable in the reduction list of the loop: initialization, partitioning of the code among processors for partial sum, and finally sum up each partial sum result. In addition, the code segment 12 of FIG. 2 illustrates the code generation for each step generated by the parallelizer.
FIG. 3 illustrates the code segment 14 for initialization. In FIG. 3, the parallelizer initializes an extreme value as a negative infinity and initializes its corresponding index with 0.
FIG. 4 illustrates the code segment 16 for partial summation of each processor in accordance with the exemplary embodiments of the present invention. In this phase, the sequential code is cloned and each reduction variable is assigned an array to hold its partial result and the index of the array is the number of this thread.
Finally, in the exemplary embodiments, the partial results are collected and summed up together by the master thread. The code segment 18 for summing up the partial results is shown in FIG. 5. FIG. 5 illustrates that the value of BUMP is always positive within the IF branch. Hence, during copy propagation pass, the code can copy the value of BUMP inside the IF branch without the max operator. Also, the loop dependency in the induction variable is eliminated through the adding and removal of the max operator to the unknown loop count within the nested loops.
The capabilities of the present invention can be implemented in software, firmware, hardware or some combination thereof.
As one example, one or more aspects of the present invention can be included in an article of manufacture (e.g., one or more computer program products) having, for instance, computer usable media. The media has embodied therein, for instance, computer readable program code means for providing and facilitating the capabilities of the present invention. The article of manufacture can be included as a part of a computer system or sold separately.
The flow diagrams depicted herein are just examples. There may be many variations to these diagrams or the steps (or operations) described therein without departing from the spirit of the invention. For instance, the steps may be performed in a differing order, or steps may be added, deleted or modified. All of these variations are considered a part of the claimed invention.
While the preferred embodiment to the invention has been described, it will be understood that those skilled in the art, both now and in the future, may make various improvements and enhancements which fall within the scope of the claims which follow. These claims should be construed to maintain the proper protection for the invention first described.